祝文斌, 苑晶, 朱书豪, 胡天帅, 高远兮, 张雪波. 低光照场景下基于序列增强的移动机器人人体检测与姿态识别[J]. 机器人, 2022, 44(3): 299-309. DOI: 10.13973/j.cnki.robot.210144
引用本文: 祝文斌, 苑晶, 朱书豪, 胡天帅, 高远兮, 张雪波. 低光照场景下基于序列增强的移动机器人人体检测与姿态识别[J]. 机器人, 2022, 44(3): 299-309. DOI: 10.13973/j.cnki.robot.210144
ZHU Wenbin, YUAN Jing, ZHU Shuhao, HU Tianshuai, GAO Yuanxi, ZHANG Xuebo. Sequence-enhancement-based Human Detection and Posture Recognition of Mobile Robots in Low Illumination Scenes[J]. ROBOT, 2022, 44(3): 299-309. DOI: 10.13973/j.cnki.robot.210144
Citation: ZHU Wenbin, YUAN Jing, ZHU Shuhao, HU Tianshuai, GAO Yuanxi, ZHANG Xuebo. Sequence-enhancement-based Human Detection and Posture Recognition of Mobile Robots in Low Illumination Scenes[J]. ROBOT, 2022, 44(3): 299-309. DOI: 10.13973/j.cnki.robot.210144

低光照场景下基于序列增强的移动机器人人体检测与姿态识别

Sequence-enhancement-based Human Detection and Posture Recognition of Mobile Robots in Low Illumination Scenes

  • 摘要: 灾难救援、地下空间开发利用等场景均存在低光照、甚至完全黑暗的问题,导致机器人目标搜索与识别困难。为此,本文面向低光照场景提出基于红外深度相机图像序列的人体检测和姿态识别方法。首先,利用基于YOLO v4的AlphaPose算法检测人体框和关键点。然后,提出基于特征点匹配的漏检人体框恢复算法,降低人体漏检率,同时使用D-S(Dempster-Shafer)证据理论融合人体框和关键点的检测结果,从而降低人体误检率。最后,设计一种基于图像序列信息的人体姿态分层识别方法,在不同的识别层提取不同的人体躯干特征,利用连续多帧躯干向量特征组成的特征序列对人体姿态进行精准的识别并进行实验验证。实验结果表明本文算法能够在低光照条件下实现准确的人体检测与姿态识别,姿态识别准确率高达95.36%。

     

    Abstract: Low illumination or even complete darkness is a very serious problem for lots of application scenes, such as disaster relief and underground space development, which brings a challenge to target search and recognition of robot. So a method for human detection and posture recognition in low illumination scenes using the image sequences collected by an infrared depth camera is proposed. Firstly, the AlphaPose algorithm based on YOLO v4 to detect human bounding boxes and key points is used. Then, a method to recover the missed human bounding box based on feature matching is proposed to reduce the missing detection rate. Meanwhile, the D-S (Dempster-Shafer) evidence theory is used to fuse the detection results of human bounding boxes and key points, in order to reduce the detection error rate. Finally, a sequence-based hierarchical recognition method to classify the human postures is designed, which extracts the torso features of human body and uses the sequential torso features in multiple frames to recognize the human posture accurately. Experimental results demonstrate that the proposed method can achieve good performance of human detection and posture recognition in low illumination scenes, and the accuracy of posture recognition can reach 95.36%.

     

/

返回文章
返回